A method and device for predicting collaborative load of multiple charging stations

Abstract

The present invention relates to the field of electric load forecasting, and particularly relates to a multi-charging station collaborative load forecasting method and device. The method and device first mine the implicit associations in the load spatio-temporal information of multiple charging stations and generate a multi-charging station schema structure from the load spatio-temporal information of multiple charging stations, convert the discrete load sequence data into graph data, and perform collaborative load forecasting for multiple charging stations, solving the drawback that traditional load forecasting models can only predict the load changes of a single charging station. Secondly, the spatio-temporal features in the load spatio-temporal information of multiple charging stations and the multi-charging station schema structure are extracted, and the extracted spatio-temporal features are used for iterative graph structure learning of the multi-charging station schema structure, so as to extract more accurate and comprehensive spatio-temporal features. Finally, a joint loss function is constructed to jointly train the process of constructing the multi-charging station schema structure and extracting spatio-temporal features, complete end-to-end load forecasting, and improve the model training and forecasting efficiency.

Description

Technical Field

[0001] The present invention relates to the field of power load forecasting, and in particular to a method and device for collaborative load forecasting of multiple charging stations. Background Art

[0002] In recent years, with the opening of the power market and the deepening of energy Internet, a large number of distributed energy devices have participated in the operation of the power market. As a special type of load, electric vehicles have random uncertainties in time, space and behavior. The disorderly access of large-scale electric vehicles to the power grid may cause a sharp increase in the load of the power grid, an increase in the peak-to-valley difference, and even cause the power grid to collapse, which brings new challenges to the safe operation of the power grid. Charging stations are the main places where electric vehicles are concentrated to access the power grid. Predicting the load changes of charging stations can effectively avoid power grid risks.

[0003] Charging station load forecasting is to predict future load changes through historical data in order to quantitatively analyze the impact of electric vehicles on the power grid. Charging station load forecasting is essentially a time series forecasting problem, that is, using load data from a period of time in the past to predict future load changes. The load forecast value of a single charging station is only related to its own historical load changes, while multiple charging stations are distributed in different geographical spaces in the city and have different spatial characteristics. Collaborative load forecasting for multiple charging stations can be regarded as a spatiotemporal forecasting problem. The load forecast value of each charging station is not only related to the historical information of the charging station itself, but also to the historical load information of other charging stations. Common methods for processing charging station load forecasting include time series modeling, regression analysis, fuzzy prediction, and deep learning methods. Among them:

[0004] (1) Time series modeling: Time series modeling is to establish a mathematical model that describes the change of power load over time based on the historical load data. On the basis of this model, the expression of load prediction is established and the future load is predicted.

[0005] (2) Regression analysis: The regression analysis prediction method is to find the correlation between the independent variable and the dependent variable and its regression equation based on the changing rules of historical data and the factors affecting load changes, determine the model parameters, and infer the load value at future times.

[0006] (3) Fuzzy prediction: Fuzzy prediction is a load forecasting technology based on fuzzy mathematics theory. The concept of fuzzy mathematics can describe some fuzzy phenomena in the power system, such as the key factors in load forecasting: the evaluation of weather conditions, the classification of load date types, etc. Applying fuzzy methods to load forecasting can better deal with the uncertainty of load changes. At present, there are mainly the following methods for applying fuzzy theory to load forecasting: fuzzy clustering method, fuzzy similarity priority ratio method and fuzzy maximum closeness method.

[0007] (4) Deep learning: The deep learning method selects the load of the past period as training samples, constructs a suitable network structure, and uses a certain training algorithm to train the network. After it meets the accuracy requirements, this neural network is used as a load forecasting model. According to different network structures, it can be divided into RNN, LSTM, GRU, MLP and other methods.

[0008] So far, in the field of charging station load forecasting, the common defects of existing technologies mainly include:

[0009] (1) Traditional mathematical statistics methods such as time series modeling, regression analysis, and fuzzy prediction rely more on experience to build models and determine parameters. They need to consider different influencing factors, resulting in complex and inaccurate modeling. At the same time, they require a large amount of different types of data for model verification, and their generalization performance is poor in different scenarios.

[0010] (2) Compared with traditional mathematical statistics methods, deep learning-based methods can automatically mine the time series characteristics in the load change sequence, with simple modeling and good results. However, the existing deep learning model only considers the load prediction of a single charging station and cannot perform collaborative prediction of multiple charging stations. Therefore, network models need to be trained separately for different charging stations, which has high computational load pressure and poor scalability. In addition, there are implicit associations between multiple charging stations. The existing model only considers the time series characteristics of the load change of a single charging station, ignoring the spatial randomness characteristics of the load distribution of multiple charging stations. The extracted load change characteristics are not accurate and comprehensive, which affects the accuracy of the prediction. Therefore, further research is needed on the mining of the spatiotemporal characteristics of the load of multiple charging stations in the city. Summary of the invention

[0011] The embodiments of the present invention provide a method and device for collaborative load prediction of multiple charging stations, so as to at least solve the technical problem that the prior art is not accurate and comprehensive enough in extracting load change characteristics.

[0012] According to an embodiment of the present invention, a method for collaborative load prediction of multiple charging stations is provided, comprising the following steps:

[0013] Mining the implicit correlation among the spatiotemporal load information of multiple charging stations and generating a multi-charging station graph structure from the spatiotemporal load information of multiple charging stations;

[0014] Extracting the spatiotemporal information of the loads of multiple charging stations and the spatiotemporal features in the multi-charging station schema structure, and using the extracted spatiotemporal features to perform iterative graph structure learning on the multi-charging station schema structure;

[0015] A joint loss function is constructed to jointly train the multi-charging station graph structure construction and spatiotemporal feature extraction processes.

[0016] Furthermore, the method further comprises:

[0017] The multi-charging station graph structure construction and spatiotemporal feature extraction process after joint training are used to predict the multi-charging station collaborative load based on the spatiotemporal information of the load of multiple charging stations.

[0018] Furthermore, before mining the implicit associations among the load spatiotemporal information of multiple charging stations and generating a multi-charging station graph structure from the load spatiotemporal information of multiple charging stations, the method further includes:

[0019] Preprocess the spatiotemporal load information of multiple charging stations, aggregate the data within each preset time period, and the aggregated data represents the total charging power of the charging station within that period;

[0020] The data were divided into three groups according to adjacent time periods, time intervals of days, and time intervals of weeks. Linear interpolation was used to fill in the missing data, and then the MinMax method was used to normalize the data.

[0021] Furthermore, the method specifically includes:

[0022] The three sets of data are input into the graph structure learning module respectively. The graph structure learning module calculates the similarity between charging stations based on the historical load change data of each charging station through a similarity measurement function, generates a multi-charging station graph structure with charging stations as nodes and similarities as edges, and inputs the graph structure information into three sets of spatiotemporal feature extraction modules;

[0023] The three sets of data and the learned graph structure information are input into three sets of spatiotemporal feature extraction modules. The spatiotemporal feature extraction modules use graph convolutional neural networks and dilated convolutional networks to extract spatiotemporal features respectively. The learned features are then input into the graph structure learning module for iterative graph structure learning. The three sets of spatiotemporal features are fused and input into the fully connected layer to predict the final result.

[0024] A loss function is designed for the graph structure learning module to control the sparsity and connectivity of the learned graph structure. A loss function is designed for the spatiotemporal feature extraction module to reduce the difference between the prediction results and the label data. The two loss functions are weightedly summed to construct a joint loss function to achieve joint training of the graph structure learning module and the spatiotemporal feature extraction module.

[0025] Furthermore, in the data input layer, three sets of data of multiple charging stations at a time interval, a time interval of days, and a time interval of weeks are input simultaneously, namely, X P ,X D ,X W ∈R M×N×D , M is the sequence length, N is the number of charging station nodes, and D is the input dimension;

[0026] In the graph structure learning module, the implicit graph structure between multiple charging station nodes is generated according to the data input. P ,X D ,X W Generate different structures G P ,G D ,G W ;

[0027] In the process of extracting spatiotemporal features, the generated graph structure is iteratively updated. In the spatiotemporal feature extraction module, it is divided into three sub-modules: ST-P, ST-D, and ST-W. The sub-modules have the same structure and the input data are X P ,X D ,X W , which are used to extract the time proximity feature, period feature and trend feature of load change respectively;

[0028] Finally, the three different spatiotemporal features are fused and the prediction results are output using the fully connected layer.

[0029] Furthermore, in the graph structure learning module, in order to learn the hidden associations between charging station nodes, the graph similarity metric learning method is adopted, and cosine similarity is designed as the metric function:

[0030] s ij =cos(w⊙x i ,w⊙x j )

[0031] where ⊙ represents the Hadamard product, w represents the learnable parameter, and x i ,x j The input data is used; the cosine similarity is extended to a multi-head version using m weight vectors, and the independent similarity matrices are calculated respectively and the average is taken as the final similarity:

[0032]

[0033]

[0034] in Calculate the input vector x i and x j The p-th cosine similarity of , each similarity is considered as part of the semantic feature of the vector;

[0035] Given a symmetric weighted adjacency matrix A of an undirected graph, the Dirichlet energy is used to measure the graph signal Smoothness:

[0036]

[0037] where tr(·) represents the trace of the matrix, L is the graph Laplacian matrix, and D = ∑ j A ij is the degree matrix; adjacent nodes have similar features by minimizing Ω(A,X); sparse constraints are added to the adjacency matrix:

[0038]

[0039] where |·|| F is the Frobenius norm, β and γ are non-negative hyperparameters; the first loss in the above formula is used to penalize the formation of a non-connected graph, and the second loss is used to penalize the node degree to control the sparsity of the graph; the smoothness loss and the sparsity constraint loss are added to obtain the overall regularization loss of the graph:

[0040] L g =αΩ(A,X)+f(A)

[0041] α is a non-negative hyperparameter; the overall regularization loss controls the smoothness, connectivity, and sparsity of the graph.

[0042] Furthermore, in the spatiotemporal feature extraction module, each ST module includes two ST sub-blocks, each ST sub-block includes two layers of gated dilated convolution and one layer of gated graph convolution, and the gated dilated convolution network is used to extract the time features of the load sequence, and the gated graph convolution is used to extract the spatial features of multiple charging stations; the gated graph convolution layer is a bridge connecting the upper and lower dilated convolution layers, and after the dilated convolution, the spatial state can be quickly propagated on the graph convolution layer; G is the learned graph structure information, which is used as a graph convolution prior in each ST sub-block, and F is the extracted spatiotemporal features, which are passed back to the IDGL graph structure learning module to iteratively learn the graph structure information;

[0043] The dilated convolution jumps a certain distance at each step. Given a one-dimensional sequence input X∈R T and a filter f∈R K , the dilated convolution at time t is expressed as:

[0044]

[0045] Where d is the dilation factor, which determines the jump distance of each convolution;

[0046] In the ST sub-block, gated dilated convolution is used to extract the timing dynamic characteristics of load changes:

[0047] H'=Dil_Conv f (X l )=f*H l

[0048] in is the input of layer l, Dil_Conv is the dilated convolution operation, is the convolution kernel, is the output, N is the number of charging station nodes, M is the length of the load change sequence, K is the convolution kernel size, C i ,C o are the number of input and output channels respectively; H' is split evenly and nonlinearity is increased using gated linear units (GLU):

[0049] (H' 1 ,H' 2 )=split(H')

[0050]

[0051] Where split represents the split operation. is the gate input, is the output, tanh and sigmoid are the activation functions; in the gated graph convolution layer, the first-order approximate GCN is embedded into the time gated unit:

[0052]

[0053] (H' 1 ,H' 2 )=split(θ*gH l )

[0054]

[0055] Where *g is the graph convolution operation, are learnable parameters, is the prior information obtained from the graph structure information G:

[0056]

[0057]

[0058] A∈R N×N is the adjacency matrix of graph G, D∈R N×N is the degree matrix; in the gated attention dilation convolution layer, the self-attention mechanism is introduced:

[0059]

[0060]

[0061] H l+1 =Dil_Conv f (Att(Q,K,V))=f*Att(Q,K,V)

[0062] Where Q, K, and V are query, key, and value matrices respectively, and w Q ,w K ,w V is a learnable parameter, Att is the attention value calculation function, H l+1 is the output of layer l.

[0063] Furthermore, the spatiotemporal features extracted by the ST-P, ST-D, and ST-W modules are Splice to The time proximity features, period features, and trend features of load changes are simultaneously input into the fully connected layer to calculate the final prediction results:

[0064]

[0065] Where W p and b p is a learnable parameter; the mean absolute error MAE is used as the prediction loss function:

[0066]

[0067] Define the joint loss function for graph structure learning and prediction:

[0068]

[0069] Among them, λ is a parameter used to balance the influence of the graph structure learning module and the spatiotemporal feature extraction module.

[0070] According to an embodiment of the present invention, a multi-charging station collaborative load prediction device is provided, comprising:

[0071] A graph structure learning module is used to mine the implicit associations among the spatiotemporal load information of multiple charging stations and generate a graph structure of multiple charging stations from the spatiotemporal load information of multiple charging stations;

[0072] A spatiotemporal feature extraction module is used to extract the spatiotemporal information of the loads of multiple charging stations and the spatiotemporal features in the multi-charging station schema structure, and use the extracted spatiotemporal features to perform iterative graph structure learning on the multi-charging station schema structure;

[0073] The joint loss function construction module is used to construct a joint loss function to jointly train the multi-charging station graph structure construction and spatiotemporal feature extraction process.

[0074] Furthermore, the device also includes:

[0075] The multi-charging station collaborative load prediction module is used to use the multi-charging station graph structure construction and spatiotemporal feature extraction process after joint training to perform multi-charging station collaborative load prediction on the spatiotemporal load information of multiple charging stations.

[0076] A storage medium stores a program file capable of implementing any of the above-mentioned multi-charging station collaborative load prediction methods.

[0077] A processor is used to run a program, wherein when the program is run, any one of the above-mentioned multi-charging station collaborative load prediction methods is executed.

[0078] The method and device for collaborative load prediction of multiple charging stations in the embodiment of the present invention first mines the implicit correlation between the load spatiotemporal information of multiple charging stations and generates a multi-charging station graph structure from the load spatiotemporal information of multiple charging stations, converts the discrete load sequence data into graph data, and performs collaborative load prediction of multiple charging stations, which solves the shortcoming that the traditional load prediction model can only predict the load changes of a single charging station. Secondly, the load spatiotemporal information of multiple charging stations and the spatiotemporal features in the graph structure of multiple charging stations are extracted, and the graph structure of the multi-charging station is iteratively learned using the extracted spatiotemporal features, thereby extracting more accurate and comprehensive spatiotemporal features. Finally, a joint loss function is constructed, and the construction of the graph structure of multiple charging stations and the spatiotemporal feature extraction process are jointly trained to construct a joint loss function to complete end-to-end load prediction and improve the efficiency of model training and prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0080] Figure 1 This is a flow chart of the method for collaborative load prediction of multiple charging stations of the present invention;

[0081] Figure 2 This is a diagram of a multi-charging station collaborative load prediction model in the present invention;

[0082] Figure 3 It is a framework diagram for iterative graph structure learning in the present invention;

[0083] Figure 4 It is a diagram of the spatiotemporal feature extraction module in the present invention;

[0084] Figure 5 It is the expanded convolutional network graph in the present invention;

[0085] Figure 6 This is a module diagram of the multi-charging station collaborative load prediction device of the present invention. DETAILED DESCRIPTION

[0086] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0087] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0088] Example 1

[0089] According to an embodiment of the present invention, a method for collaborative load prediction of multiple charging stations is provided. Figure 1 , including the following steps:

[0090] S101: mining implicit associations in the spatiotemporal load information of multiple charging stations and generating a multi-charging station graph structure from the spatiotemporal load information of multiple charging stations;

[0091] S102: extracting the spatiotemporal information of the loads of the multiple charging stations and the spatiotemporal features in the multi-charging station schematic structure, and using the extracted spatiotemporal features to perform iterative graph structure learning on the multi-charging station schematic structure;

[0092] S103: Construct a joint loss function to jointly train the multi-charging station graph structure construction and spatiotemporal feature extraction process.

[0093] The method for collaborative load prediction of multiple charging stations in the embodiment of the present invention first mines the implicit associations in the load spatiotemporal information of multiple charging stations and generates a multi-charging station graph structure from the load spatiotemporal information of multiple charging stations, converts discrete load sequence data into graph data, and performs collaborative load prediction of multiple charging stations, which solves the shortcoming that the traditional load prediction model can only predict the load changes of a single charging station. Secondly, the load spatiotemporal information of multiple charging stations and the spatiotemporal features in the graph structure of multiple charging stations are extracted, and the graph structure of the multi-charging station is iteratively learned using the extracted spatiotemporal features, thereby extracting more accurate and comprehensive spatiotemporal features. Finally, a joint loss function is constructed, and the construction of the graph structure of multiple charging stations and the spatiotemporal feature extraction process are jointly trained to construct a joint loss function to complete end-to-end load prediction and improve the efficiency of model training and prediction.

[0094] The method further includes:

[0095] The multi-charging station graph structure construction and spatiotemporal feature extraction process after joint training are used to predict the multi-charging station collaborative load based on the spatiotemporal information of the load of multiple charging stations.

[0096] Before mining the implicit associations among the load spatiotemporal information of multiple charging stations and generating a multi-charging station graph structure from the load spatiotemporal information of multiple charging stations, the method further includes:

[0097] Preprocess the spatiotemporal load information of multiple charging stations, aggregate the data within each preset time period, and the aggregated data represents the total charging power of the charging station within that period;

[0098] The data were divided into three groups according to adjacent time periods, time intervals of days, and time intervals of weeks. Linear interpolation was used to fill in the missing data, and then the MinMax method was used to normalize the data.

[0099] The method specifically includes:

[0100] The three sets of data are respectively input into the graph structure learning module 201, which calculates the similarity between charging stations based on the load history change data of each charging station through a similarity measurement function, generates a multi-charging station graph structure with charging stations as nodes and similarities as edges, and inputs the graph structure information into the three sets of spatiotemporal feature extraction modules 202;

[0101] The three sets of data and the learned graph structure information are input into three sets of spatiotemporal feature extraction modules 202. The spatiotemporal feature extraction modules 202 use graph convolutional neural networks and dilated convolutional networks to extract spatiotemporal features respectively, and then input the learned features into the graph structure learning module 201 for iterative graph structure learning, and perform feature fusion on the three sets of spatiotemporal features, and input them into the fully connected layer to predict the final result;

[0102] A loss function is designed for the graph structure learning module 201 to control the sparsity and connectivity of the learned graph structure. A loss function is designed for the spatiotemporal feature extraction module 202 to reduce the difference between the prediction result and the label data. The two parts of the loss function are weightedly summed to construct a joint loss function to achieve joint training of the graph structure learning module 201 and the spatiotemporal feature extraction module 202.

[0103] In the data input layer, three sets of data of multiple charging stations at a time interval, a time interval of days, and a time interval of weeks are input at the same time, namely, X P ,X D ,X W ∈R M×N×D , M is the sequence length, N is the number of charging station nodes, and D is the input dimension;

[0104] In the graph structure learning module 201, an implicit graph structure among multiple charging station nodes is generated according to the data input. P ,X D ,X W Generate different structures G P ,G D ,G W ;

[0105] In the process of extracting spatiotemporal features, the generated graph structure is iteratively updated. In the spatiotemporal feature extraction module 202, it is divided into three submodules: ST-P, ST-D, and ST-W. The submodules have the same structure, and the input data are X P ,X D ,X W , which are used to extract the time proximity feature, period feature and trend feature of load change respectively;

[0106] Finally, the three different spatiotemporal features are fused and the prediction results are output using the fully connected layer.

[0107] In the graph structure learning module 201, in order to learn the hidden associations between charging station nodes, a graph similarity metric learning method is adopted, and cosine similarity is designed as a metric function:

[0108] s ij =cos(w⊙x i ,w⊙x j )

[0109] where ⊙ represents the Hadamard product, w represents the learnable parameter, and x i ,x j The input data is used; the cosine similarity is extended to a multi-head version using m weight vectors, and the independent similarity matrices are calculated respectively and the average is taken as the final similarity:

[0110]

[0111]

[0112] in Calculate the input vector x i and x j The p-th cosine similarity of , each similarity is considered as part of the semantic feature of the vector;

[0113] Given a symmetric weighted adjacency matrix A of an undirected graph, the Dirichlet energy is used to measure the graph signal Smoothness:

[0114]

[0115] where tr(·) represents the trace of the matrix, L is the graph Laplacian matrix, and D = ∑ j A ij is the degree matrix; adjacent nodes have similar features by minimizing Ω(A,X); sparse constraints are added to the adjacency matrix:

[0116]

[0117] where |·|| F is the Frobenius norm, β and γ are non-negative hyperparameters; the first loss in the above formula is used to penalize the formation of a non-connected graph, and the second loss is used to penalize the node degree to control the sparsity of the graph; the smoothness loss and the sparsity constraint loss are added to obtain the overall regularization loss of the graph:

[0118] L g =αΩ(A,X)+f(A)

[0119] α is a non-negative hyperparameter; the overall regularization loss controls the smoothness, connectivity, and sparsity of the graph.

[0120] Among them, in the spatiotemporal feature extraction module 202, each ST module includes two ST sub-blocks, each ST sub-block includes two layers of gated dilated convolution and one layer of gated graph convolution, and the gated dilated convolution network is used to extract the time characteristics of the load sequence, and the gated graph convolution is used to extract the spatial characteristics of multiple charging stations; wherein the gated graph convolution layer is a bridge connecting the upper and lower dilated convolution layers, and after the dilated convolution, the spatial state is quickly propagated on the graph convolution layer; G is the learned graph structure information, which is used as a graph convolution prior in each ST sub-block, and F is the extracted spatiotemporal feature, which is returned to the IDGL graph structure learning module 201 for iterative learning of the graph structure information;

[0121] The dilated convolution jumps a certain distance at each step. Given a one-dimensional sequence input X∈R TAnd a filter f∈R K , the dilated convolution at time t is expressed as:

[0122]

[0123] Where d is the dilation factor, which determines the jump distance of each convolution;

[0124] In the ST sub-block, gated dilated convolution is used to extract the timing dynamic characteristics of load changes:

[0125] H'=Dil_Conv f (X l )=f*H l

[0126] in is the input of layer l, Dil_Conv is the dilated convolution operation, is the convolution kernel, is the output, N is the number of charging station nodes, M is the length of the load change sequence, K is the convolution kernel size, C i ,C o are the number of input and output channels respectively; H' is split evenly and nonlinearity is increased using gated linear units (GLU):

[0127] (H' 1 ,H' 2 )=split(H')

[0128]

[0129] Where split represents the split operation. is the gate input, is the output, tanh and sigmoid are the activation functions; in the gated graph convolution layer, the first-order approximate GCN is embedded into the time gated unit:

[0130]

[0131] (H' 1 ,H' 2 )=split(θ*gH l )

[0132]

[0133] Where *g is the graph convolution operation, are learnable parameters, is the prior information obtained from the graph structure information G:

[0134]

[0135]

[0136] A∈R N×N is the adjacency matrix of graph G, D∈R N×N is the degree matrix; in the gated attention dilation convolution layer, the self-attention mechanism is introduced:

[0137]

[0138]

[0139] H l+1 =Dil_Conv f (Att(Q,K,V))=f*Att(Q,K,V)

[0140] Where Q, K, and V are query, key, and value matrices respectively, and w Q ,w K ,w V is a learnable parameter, Att is the attention value calculation function, H l+1 is the output of layer l.

[0141] Among them, the spatiotemporal features extracted by ST-P, ST-D, and ST-W modules are Splice to The time proximity features, period features, and trend features of load changes are simultaneously input into the fully connected layer to calculate the final prediction results:

[0142]

[0143] Where W p and b p is a learnable parameter; the mean absolute error MAE is used as the prediction loss function:

[0144]

[0145] Define the joint loss function for graph structure learning and prediction:

[0146]

[0147] Here, λ is a parameter used to balance the influence of the graph structure learning module 201 and the spatiotemporal feature extraction module 202 .

[0148] The following is a specific embodiment, and see Figure 2-5 , the multi-charging station collaborative load prediction method of the present invention is described in detail:

[0149] In order to solve the technical problems existing in the above-mentioned prior art, the present invention designs a method for collaborative load prediction of multiple charging stations, mines the implicit associations between multiple charging stations, further utilizes the load history data, extracts the spatiotemporal characteristics of the load changes of multiple charging stations, and predicts future load changes. In the technical solution of the present invention, a graph structure learning module 201 for collaborative load prediction of multiple charging stations based on a graph convolutional neural network is proposed. The graph structure learning module 201 performs iterative graph structure learning based on the spatiotemporal information of the loads of multiple charging stations, mines the implicit associations between multiple charging stations and generates a graph structure, and converts discrete load sequence data into graph data, so that the graph convolutional neural network can be used to perform collaborative load prediction of multiple charging stations, solving the shortcoming that the traditional load prediction model can only predict the load changes of a single charging station. The present invention further designs a spatiotemporal feature extraction module 202 based on a graph convolutional neural network and a time series convolution (expanded convolution) network, learns the spatial features between multiple charging stations through a graph convolutional neural network, and uses a time series convolutional network to learn the temporal features of load changes, thereby extracting more accurate and comprehensive spatiotemporal features. In order to realize the joint training of the graph structure learning module 201 and the spatiotemporal feature extraction module 202, a joint loss function is constructed to complete end-to-end load forecasting and improve the model training and forecasting efficiency.

[0150] The process flow of the present invention is as follows Figure 2 As shown:

[0151] First, data preprocessing was performed to remove data from weekends and holidays to ensure the universality of the model. The data in each period of time was aggregated, and the aggregated data records represent the total charging power of the charging station during this period. The data was divided into three groups according to adjacent time periods, time intervals of days, and time intervals of weeks. Linear interpolation was used to fill in the missing data, and then the MinMax method was used to normalize the data.

[0152] Second, graph structure learning. The three sets of data are respectively input into the graph structure learning module 201. The graph structure learning module 201 calculates the similarity between charging stations based on the load history change data of each charging station through a similarity measurement function, and further generates a multi-charging station graph structure with charging stations as nodes and similarities as edges, and inputs the graph structure information into three sets of spatiotemporal feature extraction modules 202.

[0153] Third, spatiotemporal feature extraction, the three sets of data and the learned graph structure information are input into the three sets of spatiotemporal feature extraction modules 202, which use graph convolutional neural networks and dilated convolutional networks to extract spatiotemporal features respectively, and then input the learned feature representations into the graph structure learning module 201 for iterative graph structure learning. The three sets of spatiotemporal features are fused and input into the fully connected layer to predict the final result.

[0154] Fourth, joint model training: for the graph structure learning module 201, a loss function is designed to control the sparsity and connectivity of the learned graph structure. For the spatiotemporal feature extraction module 202, a loss function is designed to reduce the difference between the prediction result and the label data. The weighted sum of the above two loss functions is used to construct a joint loss function, thereby completing the joint training of the graph structure learning module 201 and the spatiotemporal feature extraction module 202.

[0155] The technical implementation scheme of the present invention will be described in detail below. It should be noted that the specific technical implementation scheme here does not constitute a limitation on the protection scope of the present invention.

[0156] The overall framework of the present invention is as follows Figure 2 As shown in FIG. 1 , it mainly includes two parts: a graph structure learning module 201 and a spatiotemporal feature extraction module 202. In the data input layer, three sets of data of multiple charging stations at a time interval, a time interval of days, and a time interval of weeks are input at the same time, namely, X P ,X D ,X W ∈R M×N×D , M is the sequence length, N is the number of charging station nodes, and D is the input dimension. In the graph structure learning module 201, an implicit graph structure between multiple charging station nodes is generated according to the data input. P ,X D ,X W Generate different structures G P ,G D ,G W In the process of extracting spatiotemporal features, the generated graph structure is iteratively updated. In the spatiotemporal feature extraction module 202, there are three submodules: ST-P, ST-D, and ST-W. The submodules have the same structure, but the input data are X P ,X D ,X W , which are used to extract the time proximity feature, period feature and trend feature of load change respectively. Finally, the three different spatiotemporal features are fused and the prediction results are output using the fully connected layer.

[0157] 1. Graph Structure Learning Module 201

[0158] In the graph structure learning module 201, the present invention simultaneously learns the graph structure and the graph convolutional neural network (GCN) parameters in an iterative manner. The key principle of the iterative graph structure learning framework is to use more accurate graph node feature representation to learn a better graph structure, and at the same time, extract better node features based on the learned graph structure, such as Figure 3As shown in Figure 2. Different from the method of generating graphs based on original node features, the node features learned by GCN can provide useful information for graph structure learning. On the other hand, the newly learned graph structure information can provide better graph input for GCN parameter learning.

[0159] In order to learn the hidden associations between charging station nodes, a graph similarity metric learning method is adopted, and cosine similarity is designed as the metric function:

[0160] s ij =cos(w⊙x i ,w⊙x j )

[0161] where ⊙ represents the Hadamard product, w represents the learnable parameter, and x i ,x j is the input data. To improve the expressive power, the cosine similarity is extended to a multi-head version using m weight vectors, and the independent similarity matrices are calculated respectively and the average value is taken as the final similarity:

[0162]

[0163]

[0164] in Calculate the input vector x i and x j Each similarity can be considered as part of the semantic feature of the vector.

[0165] To ensure the quality of the learned graph structure, the graph regularization method is used to control its smoothness, connectivity and sparsity. Smoothness means that the value of the graph signal changes smoothly between adjacent nodes. Given the symmetric weighted adjacency matrix A of an undirected graph, the Dirichlet energy is used to measure the graph signal. Smoothness:

[0166]

[0167] where tr(·) represents the trace of the matrix, L is the graph Laplacian matrix, and D = ∑ j A ij is the degree matrix. By minimizing Ω(A,X), adjacent nodes can have similar characteristics, thus ensuring the smoothness of the graph signal. The sparsity of the graph is controlled and sparse constraints are added to the adjacency matrix:

[0168]

[0169] where |·|| Fis the Frobenius norm, β and γ are non-negative hyperparameters. The first loss in the above formula is used to penalize the formation of a non-connected graph, and the second loss is used to penalize the node degree to control the sparsity of the graph. The smoothness loss and the sparsity constraint loss are added together to obtain the overall regularization loss of the graph:

[0170] L g =αΩ(A,X)+f(A)

[0171] α is a non-negative hyperparameter. The overall regularization loss can control the smoothness, connectivity, and sparsity of the graph, thereby ensuring the quality of the learned graph structure.

[0172] 2. Spatiotemporal feature extraction module 202

[0173] The present invention designs a spatiotemporal feature extraction module 202 based on a graph convolutional neural network and an expanded convolutional network, such as Figure 4 As shown, Figure 3 The ST-P, ST-D, and ST-W in the figure are all of the same structure. Each ST module includes two ST sub-blocks. Each ST sub-block adopts a "sandwich" structure, including two layers of gated dilated convolutions and one layer of gated graph convolutions. The gated dilated convolution network is used to extract the time features of the load sequence, and the gated graph convolution is used to extract the spatial features of multiple charging stations, thereby realizing spatiotemporal feature extraction. The gated graph convolution layer is a bridge connecting the upper and lower dilated convolution layers. After the dilated convolution, the spatial state can be quickly propagated on the graph convolution layer. The "sandwich" structure can effectively apply the "bottleneck strategy" to achieve scale compression and feature compression by performing up and down sampling before and after the gated graph convolution, thereby reducing the number of model parameters. G is the learned graph structure information, which is used as a graph convolution prior in each ST sub-block. F is the extracted spatiotemporal feature, which is passed back to the IDGL graph structure learning module 201 for iterative learning of graph structure information.

[0174] Although recurrent neural networks (RNN, GRU, LSTM, etc.) have achieved good results in time series analysis, they have problems such as time-consuming iterations and slow response to dynamic changes. The dilated convolutional network has the advantages of fast training speed, simple structure, no dependency constraints between calculation steps, and can also expand the receptive field, such as Figure 5 As shown. The dilated convolution jumps a certain distance at each step. Given a one-dimensional sequence input X∈R T and a filter f∈R K , the dilated convolution at time t can be expressed as:

[0175]

[0176] Where d is the dilation factor, which determines the jump distance of each convolution. By stacking dilated convolution layers in the form of increasing dilation factors, the receptive field of the model increases exponentially, which allows dilated convolution to extract longer time dependencies with fewer layers and reduces computational complexity.

[0177] The present invention uses gated dilated convolution in the ST sub-block to extract the timing dynamic characteristics of load changes:

[0178] H'=Dil_Conv f (X l )=f*H l

[0179] in is the input of layer l, Dil_Conv is the dilated convolution operation, is the convolution kernel, is the output, N is the number of charging station nodes, M is the length of the load change sequence, K is the convolution kernel size, C i ,C o are the number of input and output channels respectively. Split H' equally and use Gated Linear Units (GLU) to increase nonlinearity:

[0180] (H' 1 ,H' 2 )=split(H')

[0181]

[0182] Where split represents the split operation. is the gate input, is the output, tanh and sigmoid are the activation functions. In the gated graph convolution layer, in order to extract both spatial and local temporal features, the first-order approximate GCN is embedded into the temporal gating unit:

[0183]

[0184] (H' 1 ,H' 2 )=split(θ*gH l )

[0185]

[0186] Where *g is the graph convolution operation, are learnable parameters, is the prior information obtained from the graph structure information G:

[0187]

[0188]

[0189] A∈R N×N is the adjacency matrix of graph G, D∈R N×N is the degree matrix. In the gated attention dilated convolution layer, global temporal features are further extracted. At the same time, in order to capture the temporal dynamic characteristics, that is, different sequence inputs may have different temporal dependencies, a self-attention mechanism is introduced:

[0190]

[0191]

[0192] H l+1 =Dil_Conv f (Att(Q,K,V))=f*Att(Q,K,V)

[0193] Where Q, K, and V are query, key, and value matrices respectively, and w Q ,w K ,w V is a learnable parameter, Att is the attention value calculation function, H l+1 is the output of layer l.

[0194] 3. Joint loss function

[0195] The spatiotemporal features extracted by ST-P, ST-D, and ST-W modules are Splice to The time proximity features, period features, and trend features of load changes are simultaneously input into the fully connected layer to calculate the final prediction results:

[0196]

[0197] Where W p and b p is a learnable parameter. This paper uses the mean absolute error (MAE) as the prediction loss function:

[0198]

[0199] Since graph structure learning is performed iteratively during the process of spatiotemporal feature extraction, the joint loss function of graph structure learning and prediction is defined as:

[0200]

[0201] Here, λ is a parameter used to balance the influence of the graph structure learning module 201 and the spatiotemporal feature extraction module 202. By optimizing the loss function, the overall training of the model can be achieved, and the learning of model parameters can be achieved while learning the optimal graph structure.

[0202] The key points and points to be protected of the present invention mainly include the following points:

[0203] 1. The technical route of multi-charging station load prediction of the present invention is to first use graph structure learning to convert the multi-charging station load prediction problem into a graph structure data processing problem, and then use graph convolutional neural network and time series network to extract the load spatiotemporal characteristics and predict them.

[0204] 2. The multi-charging station collaborative load prediction model designed by the present invention includes an iterative graph structure learning module 201, three spatiotemporal feature extraction modules 202, and a feature fusion and prediction module.

[0205] 3. A method for iteratively mining implicit associations among multiple charging stations using the graph structure learning module 201.

[0206] 4. A method for extracting the spatiotemporal characteristics of load changes at multiple charging stations using a spatiotemporal feature extraction module 202 composed of a graph convolutional neural network and an expanded convolutional network.

[0207] 5. The joint loss function of the graph structure learning and spatiotemporal feature extraction module 202 designed by the present invention.

[0208] Compared with the existing charging station load prediction method, the advantages of the present invention are mainly reflected in:

[0209] 1. Compared with traditional mathematical statistics methods, the present invention does not need to consider too many external influencing factors, the modeling process is simple, and only a small number of model hyperparameters need to be determined during the training process, so the prediction accuracy is higher.

[0210] 2. Compared with the load forecasting method based on deep learning:

[0211] i) The present invention uses a graph structure learning method to mine the implicit associations between multiple charging stations, thereby constructing discrete charging stations into a graph structure, so that the graph convolutional neural network can be further used to simultaneously predict the load changes of multiple charging stations, breaking through the limitation that deep learning methods can only predict the load of a single charging station;

[0212] ii) The present invention not only considers the temporal characteristics of the load change of a single charging station, but also considers the spatial characteristics of the load change of multiple charging stations, so the extracted temporal and spatial characteristics are more comprehensive;

[0213] iii) When predicting the load of a new charging station, the model designed by the present invention only needs to add graph node information based on the current model, while the traditional method requires training a new model. Therefore, the model in the present invention is more scalable.

[0214] A large number of experiments have been carried out on the multi-charging station collaborative load prediction model in the present invention, and the results have proved that it is feasible and effective.

[0215] Example 2

[0216] According to an embodiment of the present invention, a multi-charging station collaborative load prediction device is provided. Figure 6 ,include:

[0217] A graph structure learning module 201 is used to mine the implicit associations among the load spatiotemporal information of multiple charging stations and generate a graph structure of multiple charging stations from the load spatiotemporal information of multiple charging stations;

[0218] A spatiotemporal feature extraction module 202 is used to extract the spatiotemporal information of the loads of multiple charging stations and the spatiotemporal features in the multi-charging station schematic structure, and use the extracted spatiotemporal features to perform iterative graph structure learning on the multi-charging station schematic structure;

[0219] The joint loss function construction module 203 is used to construct a joint loss function to jointly train the multi-charging station graph structure construction and the spatiotemporal feature extraction process.

[0220] The multi-charging station collaborative load prediction device in the embodiment of the present invention first mines the implicit correlation between the load spatiotemporal information of the multiple charging stations and generates a multi-charging station graph structure from the load spatiotemporal information of the multiple charging stations, converts the discrete load sequence data into graph data, and performs collaborative load prediction of the multiple charging stations, which solves the shortcoming that the traditional load prediction model can only predict the load changes of a single charging station. Secondly, the load spatiotemporal information of the multiple charging stations and the spatiotemporal features in the multi-charging station graph structure are extracted, and the extracted spatiotemporal features are used to iterate the graph structure learning of the multi-charging station graph structure, so as to extract more accurate and comprehensive spatiotemporal features. Finally, a joint loss function is constructed, and the multi-charging station graph structure construction and spatiotemporal feature extraction process are jointly trained to construct a joint loss function to complete end-to-end load prediction and improve the model training and prediction efficiency.

[0221] The device further comprises:

[0222] The multi-charging station collaborative load prediction module is used to use the multi-charging station graph structure construction and spatiotemporal feature extraction process after joint training to perform multi-charging station collaborative load prediction on the spatiotemporal load information of multiple charging stations.

[0223] The following is a specific embodiment, and see Figure 2-5 , the multi-charging station collaborative load prediction device of the present invention is described in detail:

[0224] In order to solve the technical problems existing in the above-mentioned prior art, the present invention designs a multi-charging station collaborative load prediction device, mines the implicit associations between multiple charging stations, further utilizes the load history data, extracts the spatiotemporal characteristics of the load changes of multiple charging stations, and predicts future load changes. In the technical solution of the present invention, a graph structure learning module 201 for collaborative load prediction of multiple charging stations based on a graph convolutional neural network is proposed. The graph structure learning module 201 performs iterative graph structure learning based on the spatiotemporal information of the loads of multiple charging stations, mines the implicit associations between multiple charging stations and generates a graph structure, and converts discrete load sequence data into graph data, so that the graph convolutional neural network can be used to perform collaborative load prediction of multiple charging stations, solving the shortcoming that the traditional load prediction model can only predict the load changes of a single charging station. The present invention further designs a spatiotemporal feature extraction module 202 based on a graph convolutional neural network and a time series convolution (expanded convolution) network, learns the spatial features between multiple charging stations through a graph convolutional neural network, and uses a time series convolutional network to learn the temporal features of load changes, thereby extracting more accurate and comprehensive spatiotemporal features. In order to realize the joint training of the graph structure learning module 201 and the spatiotemporal feature extraction module 202, a joint loss function is constructed to complete end-to-end load forecasting and improve the model training and forecasting efficiency.

[0225] The process flow of the present invention is as follows Figure 2 As shown:

[0226] First, data preprocessing was performed to remove data from weekends and holidays to ensure the universality of the model. The data in each period of time was aggregated, and the aggregated data records represent the total charging power of the charging station during this period. The data was divided into three groups according to adjacent time periods, time intervals of days, and time intervals of weeks. Linear interpolation was used to fill in the missing data, and then the MinMax method was used to normalize the data.

[0227] Second, graph structure learning. The three sets of data are respectively input into the graph structure learning module 201. The graph structure learning module 201 calculates the similarity between charging stations based on the load history change data of each charging station through a similarity measurement function, and further generates a multi-charging station graph structure with charging stations as nodes and similarities as edges, and inputs the graph structure information into three sets of spatiotemporal feature extraction modules 202.

[0228] Third, spatiotemporal feature extraction, the three sets of data and the learned graph structure information are input into the three sets of spatiotemporal feature extraction modules 202, which use graph convolutional neural networks and dilated convolutional networks to extract spatiotemporal features respectively, and then input the learned feature representations into the graph structure learning module 201 for iterative graph structure learning. The three sets of spatiotemporal features are fused and input into the fully connected layer to predict the final result.

[0229] Fourth, joint model training: for the graph structure learning module 201, a loss function is designed to control the sparsity and connectivity of the learned graph structure. For the spatiotemporal feature extraction module 202, a loss function is designed to reduce the difference between the prediction result and the label data. The weighted sum of the above two loss functions is used to construct a joint loss function, thereby completing the joint training of the graph structure learning module 201 and the spatiotemporal feature extraction module 202.

[0230] The technical implementation scheme of the present invention will be described in detail below. It should be noted that the specific technical implementation scheme here does not constitute a limitation on the protection scope of the present invention.

[0231] The overall framework of the present invention is as follows Figure 2 As shown in FIG. 1 , it mainly includes two parts: a graph structure learning module 201 and a spatiotemporal feature extraction module 202. In the data input layer, three sets of data of multiple charging stations at a time interval, a time interval of days, and a time interval of weeks are input at the same time, namely, X P ,X D ,X W ∈R M×N×D , M is the sequence length, N is the number of charging station nodes, and D is the input dimension. In the graph structure learning module 201, an implicit graph structure between multiple charging station nodes is generated according to the data input. P ,X D ,X W Generate different structures G P ,G D ,G W In the process of extracting spatiotemporal features, the generated graph structure is iteratively updated. In the spatiotemporal feature extraction module 202, there are three submodules: ST-P, ST-D, and ST-W. The submodules have the same structure, but the input data are X P ,X D ,X W , which are used to extract the time proximity feature, period feature and trend feature of load change respectively. Finally, the three different spatiotemporal features are fused and the prediction results are output using the fully connected layer.

[0232] 1. Graph Structure Learning Module 201

[0233] In the graph structure learning module 201, the present invention simultaneously learns the graph structure and the graph convolutional neural network (GCN) parameters in an iterative manner. The key principle of the iterative graph structure learning framework is to use more accurate graph node feature representation to learn a better graph structure, and at the same time, extract better node features based on the learned graph structure, such as Figure 3As shown in Figure 2. Different from the method of generating graphs based on original node features, the node features learned by GCN can provide useful information for graph structure learning. On the other hand, the newly learned graph structure information can provide better graph input for GCN parameter learning.

[0234] In order to learn the hidden associations between charging station nodes, a graph similarity metric learning method is adopted, and cosine similarity is designed as the metric function:

[0235] s ij =cos(w⊙x i ,w⊙x j )

[0236] where ⊙ represents the Hadamard product, w represents the learnable parameter, and x i ,x j is the input data. To improve the expressive power, the cosine similarity is extended to a multi-head version using m weight vectors, and the independent similarity matrices are calculated respectively and the average value is taken as the final similarity:

[0237]

[0238]

[0239] in Calculate the input vector x i and x j Each similarity can be considered as part of the semantic feature of the vector.

[0240] To ensure the quality of the learned graph structure, the graph regularization method is used to control its smoothness, connectivity and sparsity. Smoothness means that the value of the graph signal changes smoothly between adjacent nodes. Given the symmetric weighted adjacency matrix A of an undirected graph, the Dirichlet energy is used to measure the graph signal. Smoothness:

[0241]

[0242] where tr(·) represents the trace of the matrix, L is the graph Laplacian matrix, and D = ∑ j A ij is the degree matrix. By minimizing Ω(A,X), adjacent nodes can have similar characteristics, thus ensuring the smoothness of the graph signal. The sparsity of the graph is controlled and sparse constraints are added to the adjacency matrix:

[0243]

[0244] where |·|| Fis the Frobenius norm, β and γ are non-negative hyperparameters. The first loss in the above formula is used to penalize the formation of a non-connected graph, and the second loss is used to penalize the node degree to control the sparsity of the graph. The smoothness loss and the sparsity constraint loss are added together to obtain the overall regularization loss of the graph:

[0245] L g =αΩ(A,X)+f(A)

[0246] α is a non-negative hyperparameter. The overall regularization loss can control the smoothness, connectivity, and sparsity of the graph, thereby ensuring the quality of the learned graph structure.

[0247] 2. Spatiotemporal feature extraction module 202

[0248] The present invention designs a spatiotemporal feature extraction module 202 based on a graph convolutional neural network and an expanded convolutional network, such as Figure 4 As shown, Figure 3 The ST-P, ST-D, and ST-W in the figure are all of the same structure. Each ST module includes two ST sub-blocks. Each ST sub-block adopts a "sandwich" structure, including two layers of gated dilated convolutions and one layer of gated graph convolutions. The gated dilated convolution network is used to extract the time features of the load sequence, and the gated graph convolution is used to extract the spatial features of multiple charging stations, thereby realizing spatiotemporal feature extraction. The gated graph convolution layer is a bridge connecting the upper and lower dilated convolution layers. After the dilated convolution, the spatial state can be quickly propagated on the graph convolution layer. The "sandwich" structure can effectively apply the "bottleneck strategy" to achieve scale compression and feature compression by performing up and down sampling before and after the gated graph convolution, thereby reducing the number of model parameters. G is the learned graph structure information, which is used as a graph convolution prior in each ST sub-block. F is the extracted spatiotemporal feature, which is passed back to the IDGL graph structure learning module 201 for iterative learning of graph structure information.

[0249] Although recurrent neural networks (RNN, GRU, LSTM, etc.) have achieved good results in time series analysis, they have problems such as time-consuming iterations and slow response to dynamic changes. The dilated convolutional network has the advantages of fast training speed, simple structure, no dependency constraints between calculation steps, and can also expand the receptive field, such as Figure 5 As shown. The dilated convolution jumps a certain distance at each step. Given a one-dimensional sequence input X∈R T And a filter f∈R K , the dilated convolution at time t can be expressed as:

[0250]

[0251] Where d is the dilation factor, which determines the jump distance of each convolution. By stacking dilated convolution layers in the form of increasing dilation factors, the receptive field of the model increases exponentially, which allows dilated convolution to extract longer time dependencies with fewer layers and reduces computational complexity.

[0252] The present invention uses gated dilated convolution in the ST sub-block to extract the timing dynamic characteristics of load changes:

[0253] H'=Dil_Conv f (X l )=f*H l

[0254] in is the input of layer l, Dil_Conv is the dilated convolution operation, is the convolution kernel, is the output, N is the number of charging station nodes, M is the length of the load change sequence, K is the convolution kernel size, C i ,C o are the number of input and output channels respectively. Split H' equally and use Gated Linear Units (GLU) to increase nonlinearity:

[0255] (H' 1 ,H' 2 )=split(H')

[0256]

[0257] Where split represents the split operation. is the gate input, is the output, tanh and sigmoid are the activation functions. In the gated graph convolution layer, in order to extract both spatial and local temporal features, the first-order approximate GCN is embedded into the temporal gating unit:

[0258]

[0259] (H' 1 ,H' 2 )=split(θ*gH l )

[0260]

[0261] Where *g is the graph convolution operation, are learnable parameters, is the prior information obtained from the graph structure information G:

[0262]

[0263]

[0264] A∈R N×N is the adjacency matrix of graph G, D∈R N×N is the degree matrix. In the gated attention dilated convolution layer, global temporal features are further extracted. At the same time, in order to capture the temporal dynamic characteristics, that is, different sequence inputs may have different temporal dependencies, a self-attention mechanism is introduced:

[0265]

[0266]

[0267] H l+1 =Dil_Conv f (Att(Q,K,V))=f*Att(Q,K,V)

[0268] Where Q, K, and V are query, key, and value matrices respectively, and w Q ,w K ,w V is a learnable parameter, Att is the attention value calculation function, H l+1 is the output of layer l.

[0269] 3. Joint loss function

[0270] The spatiotemporal features extracted by ST-P, ST-D, and ST-W modules are Splice to The time proximity features, period features, and trend features of load changes are simultaneously input into the fully connected layer to calculate the final prediction results:

[0271]

[0272] Where W p and b p is a learnable parameter. This paper uses the mean absolute error (MAE) as the prediction loss function:

[0273]

[0274] Since graph structure learning is performed iteratively during the process of spatiotemporal feature extraction, the joint loss function of graph structure learning and prediction is defined as:

[0275]

[0276] Here, λ is a parameter used to balance the influence of the graph structure learning module 201 and the spatiotemporal feature extraction module 202. By optimizing the loss function, the overall training of the model can be achieved, and the learning of model parameters can be achieved while learning the optimal graph structure.

[0277] The key points and points to be protected of the present invention mainly include the following points:

[0278] 1. The technical route of multi-charging station load prediction of the present invention is to first use graph structure learning to convert the multi-charging station load prediction problem into a graph structure data processing problem, and then use graph convolutional neural network and time series network to extract the load spatiotemporal characteristics and predict them.

[0279] 2. The multi-charging station collaborative load prediction model designed by the present invention includes an iterative graph structure learning module 201, three spatiotemporal feature extraction modules 202, and a feature fusion and prediction module.

[0280] 3. A method for iteratively mining implicit associations among multiple charging stations using the graph structure learning module 201.

[0281] 4. A method for extracting the spatiotemporal characteristics of load changes at multiple charging stations using a spatiotemporal feature extraction module 202 composed of a graph convolutional neural network and an expanded convolutional network.

[0282] 5. The joint loss function of the graph structure learning and spatiotemporal feature extraction module 202 designed by the present invention.

[0283] Compared with the existing charging station load prediction method, the advantages of the present invention are mainly reflected in:

[0284] 1. Compared with traditional mathematical statistics methods, the present invention does not need to consider too many external influencing factors, the modeling process is simple, and only a small number of model hyperparameters need to be determined during the training process, so the prediction accuracy is higher.

[0285] 2. Compared with the load forecasting method based on deep learning:

[0286] i) The present invention uses a graph structure learning method to mine the implicit associations between multiple charging stations, thereby constructing discrete charging stations into a graph structure, so that the graph convolutional neural network can be further used to simultaneously predict the load changes of multiple charging stations, breaking through the limitation that deep learning methods can only predict the load of a single charging station;

[0287] ii) The present invention not only considers the temporal characteristics of the load change of a single charging station, but also considers the spatial characteristics of the load change of multiple charging stations, so the extracted temporal and spatial characteristics are more comprehensive;

[0288] iii) When predicting the load of a new charging station, the model designed by the present invention only needs to add graph node information based on the current model, while the traditional method requires training a new model. Therefore, the model in the present invention is more scalable.

[0289] A large number of experiments have been carried out on the multi-charging station collaborative load prediction model in the present invention, and the results have proved that it is feasible and effective.

[0290] Example 3

[0291] A storage medium stores a program file capable of implementing any of the above-mentioned multi-charging station collaborative load prediction methods.

[0292] Example 4

[0293] A processor is used to run a program, wherein when the program is run, any one of the above-mentioned multi-charging station collaborative load prediction methods is executed.

[0294] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0295] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0296] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the system embodiments described above are only schematic. For example, the division of units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0297] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed over multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0298] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0299] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0300] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for collaborative load prediction of multiple charging stations. It is characterized in that The following steps are involved: Mining the implicit associations in the spatiotemporal load information of multiple charging stations and generating an explicit graph structure of multiple charging stations; Extracting spatiotemporal features from the spatiotemporal information of the loads of multiple charging stations based on the multi-charging station graph structure, and using the extracted spatiotemporal features to iteratively learn the graph structure of the multi-charging station graph structure; Construct a joint loss function to jointly train the multi-charging station graph structure construction and spatiotemporal feature extraction process; The method further comprises: The multi-charging station collaborative load forecasting is performed using the multi-charging station graph structure construction and spatiotemporal feature extraction process after joint training based on the spatiotemporal load information of multiple charging stations; The method specifically comprises: The three sets of data are input into the graph structure learning module respectively. The graph structure learning module calculates the similarity between charging stations based on the historical load change data of each charging station through a similarity measurement function, generates a multi-charging station graph structure with charging stations as nodes and similarities as edges, and inputs the graph structure information into three sets of spatiotemporal feature extraction modules; The three sets of data and the learned graph structure information are input into three sets of spatiotemporal feature extraction modules. The spatiotemporal feature extraction modules use graph convolutional neural networks and dilated convolutional networks to extract spatiotemporal features respectively. The learned features are then input into the graph structure learning module for iterative graph structure learning. The three sets of spatiotemporal features are fused and input into the fully connected layer to predict the final result. A loss function is designed for the graph structure learning module to control the sparsity and connectivity of the learned graph structure. A loss function is designed for the spatiotemporal feature extraction module to reduce the difference between the prediction results and the label data. The two loss functions are weightedly summed to construct a joint loss function to achieve joint training of the graph structure learning module and the spatiotemporal feature extraction module.

2. The method for collaborative load prediction of multiple charging stations according to claim 1, It is characterized in that Before mining the implicit association in the load spatiotemporal information of multiple charging stations and generating a multi-charging station graph structure from the load spatiotemporal information of multiple charging stations, the method further includes: Preprocess the spatiotemporal load information of multiple charging stations, aggregate the data within each preset time period, and the aggregated data represents the total charging power of the charging station within that period; The data were divided into three groups according to adjacent time periods, time intervals of days, and time intervals of weeks. Linear interpolation was used to fill in the missing data, and then the MinMax method was used to normalize the data.

3. The method for collaborative load prediction of multiple charging stations according to claim 2, It is characterized in that In the data input layer, three sets of data of multiple charging stations at a time interval, a time interval of days, and a time interval of weeks are input simultaneously, namely , is the sequence length, is the number of charging station nodes, is the input dimension; In the graph structure learning module, the implicit graph structure between multiple charging station nodes is generated according to the data input. Generate different structures ; In the process of extracting spatiotemporal features, the generated graph structure is iteratively updated. In the spatiotemporal feature extraction module, it is divided into three sub-modules: ST-P, ST-D, and ST-W. The sub-modules have the same structure, and the input data are , which are used to extract the time proximity feature, period feature and trend feature of load change respectively; Finally, the three different spatiotemporal features are fused and the prediction results are output using the fully connected layer.

4. The method for collaborative load prediction of multiple charging stations according to claim 3, It is characterized in that In the graph structure learning module, in order to learn the hidden associations between charging station nodes, the graph similarity metric learning method is adopted, and cosine similarity is designed as the metric function: in represents the Hadamard product, represents the learnable parameters, To input data; use The weight vector extends the cosine similarity to a multi-head version, calculates independent similarity matrices separately and takes the average as the final similarity: in Calculate the input vector and No. cosine similarities, each of which is considered as part of the semantic features of the vector; Given a symmetric weighted adjacency matrix of an undirected graph , using Dirichlet energy to measure graph signals Smoothness: in represents the trace of the matrix, is the graph Laplacian matrix, is the degree matrix; by minimizing Make adjacent nodes have similar features; Add a sparse constraint to the adjacency matrix: in is the Frobenius norm, and is a non-negative hyperparameter; the first loss in the above formula is used to penalize the formation of a non-connected graph, and the second loss is used to penalize the node degree to control the sparsity of the graph; the smoothness loss and the sparsity constraint loss are added to obtain the overall regularization loss of the graph: is a non-negative hyperparameter; the overall regularization loss controls the smoothness, connectivity, and sparsity of the graph.

5. The method for collaborative load prediction of multiple charging stations according to claim 4, It is characterized in that In the spatiotemporal feature extraction module, each ST module includes two ST sub-blocks, each ST sub-block includes two layers of gated dilated convolution and one layer of gated graph convolution. The gated dilated convolution network is used to extract the time features of the load sequence, and the gated graph convolution is used to extract the spatial features of multiple charging stations. The gated graph convolution layer is a bridge connecting the upper and lower dilated convolution layers. After the dilated convolution, the spatial state can be quickly propagated on the graph convolution layer. To learn the graph structure information, it is used as a graph convolution prior in each ST sub-block. The extracted spatiotemporal features are sent back to the IDGL graph structure learning module to iteratively learn the graph structure information; The dilated convolution jumps a certain distance at each step. Given a one-dimensional sequence input and a filter , the dilated convolution at time t is expressed as: in is the dilation factor, which determines the jump distance of each convolution; In the ST sub-block, gated dilated convolution is used to extract the timing dynamic characteristics of load changes: in for Layer input, is the dilated convolution operation, is the convolution kernel, For output, is the number of charging station nodes, is the load change sequence length, is the convolution kernel size, are the number of input and output channels respectively; Average split, using gated units to add nonlinearity: in Represents a split operation, is the gate input, For output, and is the activation function; in the gated graph convolution layer, the first-order approximate GCN is embedded into the time gated unit: in is the graph convolution operation, are learnable parameters, From the graph structure information The prior information obtained: For the picture The adjacency matrix of is the degree matrix; in the gated attention dilation convolution layer, the self-attention mechanism is introduced: in are query, key, and value matrices respectively, are learnable parameters, is the attention value calculation function, for Layer output.

6. The method for collaborative load prediction of multiple charging stations according to claim 5, It is characterized in that The spatiotemporal features extracted by ST-P, ST-D, and ST-W modules are Splice to , The time proximity features, period features, and trend features of load changes are simultaneously input into the fully connected layer to calculate the final prediction results: in and is a learnable parameter; the mean absolute error MAE is used as the prediction loss function: Define the joint loss function for graph structure learning and prediction: in It is a parameter used to balance the influence of the graph structure learning module and the spatiotemporal feature extraction module.

7. A multi-charging station collaborative load prediction device, It is characterized in that include: A graph structure learning module is used to mine the implicit associations in the spatiotemporal load information of multiple charging stations and generate a graph structure of multiple charging stations from the spatiotemporal load information of multiple charging stations; A spatiotemporal feature extraction module is used to extract the spatiotemporal information of the loads of multiple charging stations and the spatiotemporal features in the multi-charging station schema structure, and use the extracted spatiotemporal features to perform iterative graph structure learning on the multi-charging station schema structure; A joint loss function construction module, which is used to construct a joint loss function and jointly train the multi-charging station graph structure construction and spatiotemporal feature extraction process; The device specifically comprises: The three sets of data are input into the graph structure learning module respectively. The graph structure learning module calculates the similarity between charging stations based on the historical load change data of each charging station through a similarity measurement function, generates a multi-charging station graph structure with charging stations as nodes and similarities as edges, and inputs the graph structure information into three sets of spatiotemporal feature extraction modules; The three sets of data and the learned graph structure information are input into three sets of spatiotemporal feature extraction modules. The spatiotemporal feature extraction modules use graph convolutional neural networks and dilated convolutional networks to extract spatiotemporal features respectively. The learned features are then input into the graph structure learning module for iterative graph structure learning. The three sets of spatiotemporal features are fused and input into the fully connected layer to predict the final result. A loss function is designed for the graph structure learning module to control the sparsity and connectivity of the learned graph structure. A loss function is designed for the spatiotemporal feature extraction module to reduce the difference between the prediction results and the label data. The two loss functions are weightedly summed to construct a joint loss function to achieve joint training of the graph structure learning module and the spatiotemporal feature extraction module.

8. The multi-charging station collaborative load prediction device according to claim 7, It is characterized in that The device also includes: The multi-charging station collaborative load prediction module is used to use the multi-charging station graph structure construction and spatiotemporal feature extraction process after joint training to perform multi-charging station collaborative load prediction on the spatiotemporal load information of multiple charging stations.

Images